AI Adoption Opportunity · Project 5

Is AI worth it
for our supply chain?

The SCM Analytics Cockpit — a working answer for Cleo, CEO of a ~5,000-person cloud / hosting enterprise. Not opinions: a live system you can click.

5,000+

Employees

€640m

Managed spend

~85%

Forecast accuracy · backtested

▶ live cockpit ⌥ github repo

Step 01/ Use-case discovery selected

The problem worth solving

Sector & size: cloud / data-center infrastructure (DACH) · ~5,000 staff · €640m spend.
Stakeholders: procurement, capacity planners, finance, ops — all hit when forecasts miss.
Four pains: forecasts are wrong · deliveries slip · too much manual work · too many people in the loop.
Use case chosen: AI demand forecasting + a dynamic reorder point — strongest evidence, lowest lift.

proof →research/use_case_discovery.md

Step 02/ Market research → our play evidence-backed

The market proves it — here's our play

Market signal

AI is mainstream: 78% → 88% of orgs use AI in ≥1 function.
Supply chain is a proven savings area — 61% report savings; forecasting cuts error 30–50%.
Why now: ~90% of advanced chips from Taiwan; memory ~4× in 2025 → static reorder points break.

What AI

ML demand model + reliability scoring, with Claude only to explain it.

Where

Bolted onto demand planning → reorder → approval. Smarter, not a new system.

How it helps

Fewer stockouts, less overstock; reorder point moves itself — fewer people firefighting.

How it grows

One category → all SKUs → from advising toward acting as trust is earned.

proof →market_research.mdimplementation_plan.mdsources.md

Step 03/ Hype vs evidence

What's real, what's noise

Supported by data

AI forecasting cuts error 30–50% — treated as an upper bound, not a promise.
Forecasting hardware is genuinely hard for classical methods — structural, not marketing.

Hype / must validate

Only 39% see real EBIT impact; ~64% of projects stall in pilot.
Most supply-chain savings are <10% — our base case, not the vendor 30–50%.

proof →research/hype_vs_evidence.md

Step 04/ Opportunity & risk map

Where the upside — and the trap — live

Opportunities

Demand forecasting + dynamic reorder — prioritised: best evidence, working PoC.
Should-cost / margin lever & TCO views — secondary, already modelled.

Top risk

Model drift — the #1 reason AI fails to scale. Accuracy decays as conditions shift.
Mitigation: a named model owner + retrain cadence + accuracy SLA.

proof →research/opportunities_risks.md

Step 05/ The AI — which, and why 1 / 2

We considered deep learning — and said no

Why deep learning is the wrong tool here

Too little data: a few thousand SKUs, ~12 months, lots of zeros → it overfits and invents patterns.
Black box: it can't tell a CEO why — that breaks our "every number drills to its formula" promise.
Overkill & drifts harder: GPU training and MLOps to look impressive, then harder to monitor.

Layer 1 · forecasting (ML)

Right-sized, explainable methods for intermittent demand, scored by WMAPE / bias / drift. Sees outside drivers ARIMA can't.

Layer 2 · the LLM (advisory)

Claude (Anthropic), one click. Explains "why was it wrong / what to do" — and never invents a figure.

Step 05/ Data + how code and AI interact 2 / 2

What it predicts from · where it lives

Inputs: demand history (planned vs actual volume), PO timing & lateness, lead-time / disruption signals.
Where: not in the dashboard — in the SCM Master backend (FastAPI + DB on Railway). The dashboard is a read-only proxy.

The code ↔ AI handoff

Code fetches the raw numbers from the API.
Code computes the facts — deterministic rules engine, free, every refresh.
On a click, code hands the finished findings to Claude (structured, not raw data).
Claude narrates over them — never re-derives a number.

The AI is downstream of the math — never upstream.

proof →dashboard_documentation.mddeploy/insights.js

Step 06/ The dashboard

One source of truth, two front-ends

Built & deployed

Power BI report (the lab deliverable) + a live web cockpit.
Both read the same live API → the numbers always agree.

5–7 CEO metrics

Forecast Accuracy (1−WMAPE) · Total Spend · Top-Supplier Share
Stockout Risk · Delivery Accuracy % · Addressable Savings · TCO

proof →Order_Accuracy_Forecast_2026.pbixdashboard_documentation.md

Step 07/ Recommendation + plan pilot

🟡 Run a 10-week pilot

Not invest-at-scale. Not wait. The direction is evidence-backed; the magnitude is unproven on our SKUs.

Flips to "invest" if: a holdout backtest hits the 30–50% band and a model owner is named.
Flips to "stop" if: pilot is no better than the current spreadsheet.

validate→ data prep→ PoC · beat baseline→ pilot→ rollout + owner→ monitor

proof →solution_proposal.mdimplementation_plan.mdtimeline_estimate.md

Step 08/ What it costs to run

No team. No licence trap. Just infra.

Line	Pilot	Rollout · ~15 users
Claude tokens	~€5–20/mo	~€20–70/mo
Hosting + DB (Railway: FastAPI + Node + Postgres)	~€20–50/mo	~€30–80/mo
Power BI Pro seats (~€10/seat)	~€30/mo (3)	~€150/mo (15)
Forecast retrain compute (Prophet/statsforecast, CPU)	~€0	~€0–50/mo
Total vendor cash	~€60–100/mo	~€2.5–4k/yr

The AI tokens are the smallest line on the bill — the data refresh costs 0 tokens; the LLM only fires on a deliberate click.

proof →cost_estimation/cost_analysis.md

● Deliverables map complete

Every rubric box → a file

Use-case discovery	research/use_case_discovery.md	✓
Market research	research/market_research.md	✓
Opportunity / risk map	research/opportunities_risks.md	✓
Hype vs evidence	research/hype_vs_evidence.md	✓
Sources	sources.md	✓
Data + requirements	data/ · requirements.txt	✓

Dashboard (.pbix)	Order_Accuracy_Forecast_2026.pbix	✓
Dashboard docs	dashboard/dashboard_documentation.md	✓
Solution proposal	implementation/solution_proposal.md	✓
Implementation plan	implementation/implementation_plan.md	✓
Cost / timeline	cost_estimation/	✓
Live web cockpit	deploy/ → Railway	✓

▶ Live demo ~2 min

Let's open the cockpit

▶ scm-power-bi-production.up.railway.app

3D Control Tower — "every count is live data, not animation."
Overview → Forecast Accuracy ~85%, then Predicted vs Actual (the money shot).
Click-to-drill a KPI → the formula slide-out — "no black boxes."
Forecast tab → worst category (Networking ~21%) — "we show our weakest number."
✨ AI commentary once — "this is Layer 2: it explains, it doesn't invent."

Is AI worth itfor our supply chain?

The problem worth solving

The market proves it — here's our play

What's real, what's noise

Where the upside — and the trap — live

We considered deep learning — and said no

One source of truth, two front-ends

No team. No licence trap. Just infra.

Every rubric box → a file

Let's open the cockpit

Controls

Is AI worth it
for our supply chain?